RU2297662C2 - Method for high speed control over blocks for instant copying in data storage systems with joint usage of memory by n units - Google Patents

Abstract

FIELD: engineering of computers for controlling memory, in particular, external memory controllers.

SUBSTANCE: memory control device for operation in memory controller network contains memory controller being an owner unit, capable of controlling the blocking of certain data area during execution of input-output outputs, and component for exchanging messages, providing for transmission of at least one message with blocking request, permission of blocking, blocking removal request and blocking removal signal, and also input-output component, while any image of aforementioned data area, received by instant copying thereof, is maintained as coherent relatively to data area itself, and input-output component may position previous direct confirmation, that this data area remains coherent to any such image, to cash-memory, and may perform input-output operations on basis of aforementioned previous direct confirmation. Method describes operation of aforementioned device. Software product for computer is realized on machine-readable carrier and contains a program recorded thereon, realizing operations of aforementioned method.

EFFECT: expanded functional capabilities.

3 cl, 3 dwg

Description

FIELD OF THE INVENTION

The present invention relates to computing devices for managing memory, in particular to external memory controllers with advanced functions in storage systems with shared memory n nodes, with the function of instant copy data.

State of the art

In the field of computer storage systems, the so-called "advanced functions" are becoming more and more popular. These functions go beyond the simple I / O control implemented by conventional memory controller systems. Advanced functions are widely known and rely on the management of metadata, which are used to store information about the state of real or "user" data stored in the system. Advanced functions allow you to quickly perform various actions with virtual images, or "snapshots" of data, while real data remains available for use by user applications. One such well-known advanced feature is instant, or instant, copying of data.

At the highest level, instant copying is a function that creates a secondary image of some data. Such a function is sometimes called a copy of data at a control point, or a T0 copy. The contents of the secondary image are initially identical to the contents of the primary. The secondary image is created "instantly". In practice, this means that the secondary image is created in much less time than would be required to create a real separate physical copy, and that there is no undesirable disruption to the programs using the data being copied.

The secondary copy created in this way can be used for various purposes, including backing up, checking the system, and extracting information from data. In this case, the primary copy continues to be used by the original application for the original purposes. In contrast to this function, to create a backup without instant copy, you must first close the application, make a backup, and only then you can start the application again. Finding time windows when the application in use can be closed without compromising the workflow is becoming increasingly difficult. Thus, backup costs increase. For enterprises, the ability to create backups using the function of instant copying data without stopping the workflow is of significant and increasing value.

In solutions that implement the instant copy function, the illusion of the existence of a certain data image is created, which is achieved by redirecting read requests addressed to the secondary data image (hereinafter referred to as the destination), to the original image (hereinafter referred to as the source), unless writing is made to this data area. When a record has to be made to a specific data area (to the source or destination), in order to preserve the illusion that both the source and the destination have their own copy of the data, the process of pausing the write command is initiated and, until the recording has taken place, a read-write command is issued areas of data from the source, the read data is written to the destination, and then (and only if all the steps are successfully completed) the blocking of the write operation is released. With subsequent write commands to the same area, you can not pause their execution, since the destination already has its own copy of the data. This recording copy technique is well known and used in many operating environments.

The basis of all decisions that implement the instant copy function is the use of a data structure that controls the adoption of the decisions described above, namely, decisions as to whether read requests received by the addressee are sent to the source or destination, and whether it is necessary to suspend the write operation for copying when recording. The aforementioned data structure essentially keeps track of areas or pieces of data copied from source to destination, as opposed to data that has not been copied.

Maintaining such a data structure (hereinafter referred to as metadata) is the key to the implementation of the algorithm underlying instant copying.

Instant copying is relatively simple to implement in one processor complex (possibly with symmetric multiprocessor modules), as is often used in modern memory controllers. At the cost of a few great efforts, you can implement fault-tolerant instant copying when two (at least two) processor complexes have access to a copy of metadata. In case of failure of the first complex, the second complex can be used to continue operation without loss of access to the addressee.

However, the performance of one processor complex for performing I / O operations is limited. Even with an increase in the performance of one processor complex, expressed in the number of I / O operations per second or in bandwidth (Mbps), it has a finite limit, which ultimately imposes a limitation on system performance when working with applications. Such a limit exists in many instances of the use of instant copy technology, but the most obvious example is the memory controller. A typical memory controller has one processor complex (or perhaps a couple of processor complexes duplicating each other), which determines the performance limit of such a controller.

You can install additional memory controllers. But such separate memory controllers do not share metadata and, therefore, do not interact when managing instant data. The amount of memory becomes fragmented, and the scope of the instant copy function is limited to the system of one controller. The source disk and destination disk must be under the control of the same memory controller. The disk space of one such memory controller may be full, while the disk of another controller may have free space, but the source disk and the destination disk cannot be divided by transferring the destination disk under the control of another controller. (This is especially annoying in the case of a new snapshot, when moving the destination does not require significant resources, because no physical data is associated with the destination.)

In addition to limiting the potential effectiveness of the source-destination pair, the limited functions of the data storage system with a single controller further complicate the administration of data storage facilities.

Modern memory management systems usually do not attempt to solve this problem. These systems implement instant copy methods that allow the use of only one controller, and therefore their functionality is limited by the performance of such a controller.

An easy way to enable multiple controllers to share the instant copy function is to use one of the controllers as the owner of the metadata, while the rest of the controllers should send them all read and write requests. Using the algorithm described above, the controller owning the data processes the I / O requests as if they were coming directly from its host servers assigned to it, and returns each completed I / O request to the controller that sent it.

The main drawback of such a system and the reason that it is spreading are the fact that the cost of sending each I / O request is so high that they can even double the cost of resources at the system level and, therefore, reduce the system performance by about half.

It is known, for example, in the field of distributed database systems with parallel data processing, the use of a distributed locking control structure that uses a two-phase locking protocol to lock data in order to ensure the coherence of any data copies. However, two-phase blocking is usually time-consuming and leads to a decrease in system performance due to the need for messaging. The original two-phase blocking protocol, as such, in its original form, is not practical to use in lower-level systems of a hardware-software complex, such as storage networks with distributed memory controllers, in which the influence of the passage of blocking control messages on network performance is even more important than at the database management level.

Thus, it is desirable to take advantage of the distributed control of locks in computing environments that implement the instant copy function, while minimizing the cost of computing resources for transmitting blocking messages.

Summary of the invention

To solve the aforementioned problem, the present invention proposes a memory management device for operating in a network of memory controllers, including a memory controller, which is an owner node capable of controlling the blocking of a certain data area during I / O operations, and a messaging component that provides transmission of at least at least one message with a request for blocking, blocking permission, request for unlocking and a signal for unlocking, containing an input-output component, which carries out I / O operations with respect to data belonging to any owner node, provided that this component complies with the blocking protocols managed by this owner node, and any image of the specified data region obtained by instantly copying the latter is maintained coherent with respect to the region itself data, and the I / O component can cache the previous direct confirmation that this data area has remained coherent to any such image and perform operations ode-output based on this previous direct confirmation.

Preferably, the I / O component can discard the direct cache coherence placed in the cache and subsequently request a lock again.

By such selective discarding of the cache direct confirmation of coherence, a limited portion of the cache can be controlled.

It is also preferred that a direct confirmation of coherence with respect to one data area is also a direct confirmation of coherence with respect to another adjacent data area.

The object of the present invention is also a method of managing memory in a network of memory controllers, including a memory controller, which is an owner node that can control the blocking of a certain data area during I / O operations, and a messaging component that transmits at least one request message blocking, blocking permission, request for unlocking and signal for unlocking, which consists in the fact that I / O operations are performed in relation to data, belonging to any owner node, provided that the blocking protocols managed by this owner node comply, they support the image of the data region obtained by instantly copying the last one, coherent with respect to the data region itself, and place the previous direct confirmation that the specified the data area remains coherent with any such image, with I / O operations based on this previous direct confirmation.

When implementing the method of the invention, it is preferable to discard the direct confirmation of coherence placed in the cache memory, followed by a repeated blocking request.

In this case, by selectively dropping the cache of direct confirmation of coherence, a limited portion of the cache can be controlled.

It is also preferred that a direct confirmation of coherence with respect to one data area is also a direct confirmation of coherence with respect to another adjacent data area.

Another object of the invention is a computer software product, materially implemented on a computer-readable medium and containing a program that, when loaded into a computer system and executed in it, controls a memory management device in a network of memory controllers, which includes a controller that is the data owner node that controls locking a data area during I / O, and a messaging component that provides at least one message requesting a lock , Interlocking resolution request unlock signal and the lock release, with enforcement of the memory management unit following actions: performing input-output operations in relation to data belonging to the owner of any node, subject to compliance with the locking protocols controlled by the node-holder; maintaining the image of the data region obtained by instant copying of the latter, coherent with respect to the data region itself; placing in the cache the previous direct confirmation that the specified data area remains coherent to any such image, and I / O operations are performed based on this previous direct confirmation.

In a preferred embodiment of the present invention, a two-phase blocking messaging scheme is used to coordinate the operation of several memory controllers (or nodes) in a system with n nodes (n-sided system). Message passing ensures coordination of the system nodes, but each node is still responsible for its own I / O operations. Each node of the network has a cache facility for caching the results of its previous blocking requests into the cache, which allows this node to reuse some of the results, avoiding requests regarding the state of a particular data area, if such a state is already directly reflected by the previous results placed in cache memory.

Brief Description of the Drawings

The following describes, by way of example, a preferred embodiment of the present invention with reference to the accompanying drawings, in which:

figure 1 is a block diagram illustrating an embodiment of a two-phase blocking scheme using blocking messages to control the coherence of the data region and its image obtained by instant copying,

figure 2 - components proposed in the invention of the system in its preferred embodiment,

figure 3 - additional operations performed during the implementation of the invention in its preferred embodiment.

Detailed Description of a Preferred Embodiment

For a better understanding of the preferred embodiment of the present invention, consideration should be given to two-phase blocking messaging in order to coordinate the operation of several memory controllers (or nodes) in a data storage system having n nodes (n-sided system).

As an example, consider a system with n nodes that implements the instant copy function of a data region. Assume that each node has access to memory controlled by a group of n interacting nodes. At step 102, one of the nodes is designated as the owner of the metadata relating to all the I / O relationships in a particular data area. The remaining nodes are assigned by clients. In addition, in the currently most preferred embodiment of the invention, one of the client nodes is designated as the backup owner and maintains a copy of the metadata, which in case of failure of the owner node ensures their constant availability.

Consider the case when the I / O request, which at step 104 is received from the host system (from the English "host" - in this case, a stand-alone computer, such as a server or workstation) to a specific client node C. It is assumed that an I / O request from the host system refers to reading or writing to the destination disk, or possibly by writing to the source disk. The client node C begins processing the data by pausing the I / O requests in step 106. Then, in step 108, the client node C sends a REQ message to the owner node O asking if the corresponding fragment of the data area has been copied.

After receiving the REQ request, the O node owns the structure of its own metadata. If the owner node O determines that the data area has already been copied, in step 110, the owner node O transmits a NACK (negative acknowledgment) response. If the owner node O determines that the data region has not yet been copied, it will place a lock record in front of the corresponding metadata related to the above data region and located in its own metadata structures, and in step 112 it will respond to the client node by sending it a GNT message ( lock request completed). A lock record is necessary to ensure compatibility between a request that has just been received and executed and subsequently received requests that may affect the same metadata while processing data in client node C. Continue to maintain the lock record and the compatibility restrictions are similar to the procedure and restrictions in that case, if the I / O request was received locally by node O, which is well known to specialists in this field of technology.

After receiving the NACK message, the client node C in step 114 resumes the execution of the suspended initial I / O request.

After receiving the GNT message, the client node C continues to work in step 116, performing one or more data transfer operations required by the instant copy algorithm. In the case of a read operation from the destination, this means that the reading is applied to the source disk. Some time later, at step 118, the client node C completes the processing of the read request and, at step 120, transmits the UNL message (unlock lock) to the O node at the same time as sending the message on completion of the processing of the initial I / O request to the host system that sent such a request.

After receiving the UNL message, the O host in step 122 removes the lock record from its metadata table, thereby resuming processing of other I / O requests that were suspended due to such a lock. In the most preferred embodiment of the invention under consideration, in step 124, the owner node O transmits the UNLD message (lock is released) to the client node C, allowing the client node C to reuse the resources indicated in the original request. However, the instant copy algorithm itself does not require this.

In the case of a write operation (to the destination or source), the client node C must perform a copy operation during recording at step 127. After completing all the copying steps during recording and while the initial I / O write request remains suspended, at step 126, the client node C transmits to the node O a UNLC message (request to unlock the copy).

After receiving the UNLC message, the O owner node in step 128 marks the corresponding area in the metadata as copied, removes the write lock in step 130, responds to all pending requests in step 132, notifying that the data area is copied, and then passes to the node in step 134 client C is an UNLD message.

After receiving the UNLD message, the client node C in step 136 resumes the paused write operation, which is performed after a while, and then in step 138 informs the host system of the completion of the write operation.

In the event of an I / O error on the disk, a messaging system failure, or a node failure, data recovery paths are necessary, however, the requirements for such paths and their implementation are well known in the art.

The operations described above are described with respect to one I / O request and from the point of view of one client node C. However, the operation of the described circuit does not raise questions when there are many I / O requests coming from many client nodes that the owner node O processes, using the algorithm described above.

Figure 2 shows the device of the invention in the currently preferred embodiment, implemented in a network of memory controllers, including an owner node 202, a client node 204, an input / output component, a metadata section 206 related to data 208, which is controlled by a network of memory controllers, a copy 209 of data 208 and communications. The device has a component 210 that controls the distribution of data ownership rights and secures the owner of the metadata to the node 202, and a lock management component 212 that can control the lock at the metadata level 206 during I / O to ensure coherence with any copy 209. The device also includes a component 214 for messaging, which is located at the host node 202 and serves to exchange between the client node 204 and the host node 202 one or more messages with a request Som metadata condition, the lock setting resolution request unlock and unlock signal. The client node 204 performs I / O operations on data whose metadata is owned by the owner node 202, provided that the client node 204 performs blocking protocols at the metadata level managed by such owner node 202.

The system and method described above allow distributed control of locks in the n-shared network of controllers of shared memory, however, the disadvantage of this system is the significant overhead associated with messaging in the system. In systems with a relatively small number of controllers or relatively low activity, this drawback is not so significant, which, however, cannot be said about modern data storage systems, such as large networks of storage devices with a large number of controllers and very high memory access activity. In such circumstances, it is advisable to eliminate excessive messaging.

Therefore, to improve the process of using the data processing system in the most preferred embodiment of the present invention, it is possible for each client node to store information in which the last response received from the owner node is recorded. In particular (see FIG. 3, supplementing the block diagram in FIG. 1), the client node C is allowed at step 308 to cache information on its receipt of the NACK message after step 114 shown in FIG. 1, or transmitting and acknowledging to him a pair of UNLC / UNLD messages in step 126 and after step 134 shown in FIG.

After receiving an I / O request from the host system at step 302, similar to the step 104 shown in FIG. 1, the client node C in this case uses the modified lock control algorithm described below.

First, at step 303, the client node C checks to see if its cached data contains an indication that the affected data area has already been copied. If so, then at step 304, the client node C continues processing the I / O request without transmitting protocol messages to the node O.

If the cache does not contain such an indication, the protocol described above is used without modification. The client node C performs step 106 and the subsequent steps shown in FIG. If you receive a NACK message or a pair of UNLC / UNLD messages in step 306, the information in the cache is updated in step 308, and after such information is found in the cache in step 303, the processing of subsequent I / O requests that affect such an area may continue at step 304 without transmitting protocol messages.

When describing the technology necessary for the operation of the present invention in its currently most preferred embodiment, the term "pessimistic cache" is sometimes used. This means that the client node may not have a full updated copy of the metadata of the owner node: the client node may consider that a certain memory area needs to be copied, and the owner node will correct it by notifying the NACK message that copying is not required. But the client node should never assume that a certain area is copied unless the owner node knows the opposite.

For the system to function correctly when caching lock information in accordance with the most preferred embodiment of the invention, a number of changes are required in the operation of the client node. Firstly, each time you start the instant snapshot interaction (in step 300a) in step 301, you need to initialize the cache (to indicate that all areas need to be copied). This can be done in various ways, but the simplest is to tell the owner node to the client node. Secondly, every time a client node in step 300b could skip a message that the cache was re-initialized (possibly due to a violation of the mode in the power system), the client node should proceed from the worst case and in step 301 Re-initialize your cache or double-check the correctness of the data contained in it.

Further additions and changes obvious to those skilled in the art can be made to the method of the invention. For example, cached information can be discarded, i.e. do not take into account, since it can always be restored using the owner node, in which the only and truly updated copy is stored. Thus, in the client node for caching information, a smaller part of the space allocated for metadata can be allocated than would be necessary to store all metadata available in all nodes of the network. In this case, the client nodes could rely on the location of access with respect to the I / O operations that they process, in order to be able to continue to use the caching of the information contained in the blocking messages.

In another embodiment of the invention with enhanced functionality, a NACK message (as well as GNT or UNLD messages) may contain additional information beyond that directly related to the area processed using REQ / GNT / UNLC / UNLD messages. Host nodes can also forward client nodes information regarding neighboring areas that have also been cleared.

A typical embodiment of the above method is implementation in software running on one or more processors (not shown in the drawings), and this software may be a computer program element recorded on a suitable storage medium (also not shown), such as magnetic or optical disk. Similarly, data transmission channels may include any kind of information storage medium, as well as signal transmission media such as wired or wireless signal transmission channels.

The present invention may suitably be implemented as a software product for use in computing systems. The invention can be implemented in the form of a sequence of computer-readable instructions recorded on a tangible medium, such as a computer-readable medium, such as a floppy disk, ROM on a compact disc (CD-ROM), ROM or hard disk, or transmitted to a computing system via a modem or other interface device with using a material medium, including (and not only) optical or analog communication lines, or an intangible way using wireless technologies, including (and not only) microwave equipment, infrared or other signal transmission technology. In the sequence of machine-readable commands, the functionality described above is partially or fully implemented.

For specialists, it is clear that such machine-readable instructions can be written in a number of programming languages applicable in a variety of computer architectures or operating systems. In addition, such commands can be stored using any storage technology, existing or future, including (and not only) semiconductor, magnetic or optical technology, or transmitted using any communication technology, existing or future, including (and not only) optical, infrared or microwave technology. It is assumed that such a computer software product can be distributed in the form of removable media with accompanying documentation in printed or electronic form, for example, in the form of closed software (software package without revealing the internal structure) preloaded into a computer system, for example, on a system ROM or a non-removable disk, or distributed from a server or electronic bulletin board over a network, for example, via the Internet or the World Wide Web.

It will be apparent to those skilled in the art that various changes may be made to the above-described embodiment of the present invention.

Claims (9)

1. A memory management device for operating in a network of memory controllers, including a memory controller, which is the owner node, capable of controlling the blocking of a certain data area during I / O operations, and a messaging component that provides at least one blocking request message , a blocking permission, an unlocking request, and an unlocking signal, comprising an input / output component for performing input / output operations on data belonging to to any owner node, provided that this component matches the blocking protocols managed by this owner node, and any image of the specified data region obtained by instantly copying the latter is maintained coherent with respect to the data region itself, and the input / output component can be placed in cache previous direct confirmation that this data area has remained coherent to any such image, and perform I / O based on this previous direct confirmation. 2. The device according to claim 1, in which the input-output component can discard the direct confirmation of coherence placed in the cache memory and subsequently again request a lock. 3. The device according to claim 2, in which, by selectively discarding a direct confirmation of the coherence placed in the cache memory, a limited section of the cache memory is controlled. 4. The device according to claim 1, in which a direct confirmation of coherence with respect to one data area is also a direct confirmation of coherence with respect to another adjacent data area. 5. A method of managing memory in a network of memory controllers, including a memory controller, which is the owner node, capable of controlling the blocking of a certain data area during I / O operations, and a messaging component that provides the transmission of at least one message with a lock request, resolution blocking, a request for unlocking and a signal for unlocking, namely, that perform I / O operations with respect to data belonging to any owner node, provided that the blocking protocols managed by this owner node are consistent, support the image of the data region obtained by instantly copying the latter, coherent with respect to the data region itself, and put the previous direct confirmation that the indicated data area remains coherent to any such image in the cache, moreover, I / O operations are based on this previous direct confirmation. 6. The method according to claim 5, in which the direct confirmation of the coherence placed in the cache is discarded, followed by a repeated lock request. 7. The method according to claim 6, in which, by selectively discarding a direct confirmation of the coherence placed in the cache memory, a limited portion of the cache memory is controlled. 8. The method according to claim 5, in which a direct confirmation of coherence with respect to one data area is also a direct confirmation of coherence with respect to another adjacent data area. 9. A computer software product materially implemented on a computer-readable medium and containing a program that, when loaded into a computer system and executed in it, controls a memory management device in a network of memory controllers, which includes a controller that is the data owner that controls the blocking of the data area when performing I / O operations, and a component for messaging that provides the transmission of at least one message with a lock request, lock permission, request to remove Ia lock and unlock the signal, with enforcement of the memory management unit following: performing I / O operations with respect to data belonging to any owner node, provided that the blocking protocols managed by this owner node are consistent, maintaining the image of the data region obtained by instant copying of the latter, coherent with respect to the data region itself, placing in the cache the previous direct confirmation that the specified data area remains coherent to any such image, and I / O operations are performed based on this previous direct confirmation.

Images